Questions about dataset collection

[w4nanch1]

Hi, thank you for releasing such wonderful work and the clean codebase!

I have a question about how you collect and preprocess the data for training the action VAE, especially for the π0 experiments.

1. Episode selection from DROID / pre-training datasets

   For π0, the paper mentions using around 3000 episodes from Open-X Embodiment and DROID as the pre-training dataset for the action VAE.
   Could you share a bit more detail on how these episodes are selected in practice?

   • Do you sample episodes uniformly at random from the full DROID (and other) datasets?
   • Is there any filtering based on task type, trajectory length, success / failure labels, or embodiment?

2. Sub-trajectory splitting based on action chunks

   After you collect the episodes and load the qpos sequences (shape (T, action_dim)), how do you construct the sub-trajectories (action chunks) used as VAE training samples?

   My current understanding is that you take a sliding window over the action sequence with a fixed action chunk length s_length (e.g., s_length = 50 for π0), and use these chunks as VAE inputs.

   Could you confirm whether it works like this, or correct me if I’m misunderstanding:

   • Suppose an episode has actions a_0, a_1, ..., a_{T-1}.
   • We choose an action chunk length L = s_length (e.g., L = 50) and a stride stride (e.g., stride = 5).
   • Then we construct sub-trajectories like:
     • {a_0, ..., a_49},
     • {a_5, ..., a_54},
     • {a_10, ..., a_59},
     • and so on, until we reach the end of the episode.

Thanks again for your great work!